Accurate hypoxic surfacing detection is crucial for early warning of abnormal behaviors in industrial Recirculating Aquaculture Systems (RAS), Q1-2: as this phenomenon shows a strong correlation with Dissolved Oxygen (DO) concentration fluctuations and provides direct behavioral evidence of hypoxic stress. However, reliable detection remains difficult under complex aquatic conditions, where image degradation, background interference, feature overlap, and the small-object jointly reduce detection accuracy. To address these challenges, this study proposes MSYOLO, an enhanced hypoxic surfacing fish detection framework based on YOLOv10. The proposed model improves multi-scale representation by introducing multi-scale convolutions into both the Conv and C2f modules, strengthens salient feature extraction through the integration of a Squeeze-and-Excitation mechanism in the backbone, and enhances small-target perception by adding a P2 detection layer. Experimental results demonstrate that MSYOLO achieves precision, recall, F1-score, and mAP@0.5 of 88.6%, 89.8%, 89.3%, and 93.2%, respectively, outperforming the baseline YOLOv10n model by 2.8%, 2.6%, 2.9%, and 3.0%. Ablation experiments further verify that the proposed modules contribute consistently to performance improvement. In addition, the model maintains a favorable balance between detection accuracy and computational efficiency, which supports practical edge deployment in RAS. These findings indicate that MSYOLO not only provides an effective solution for hypoxic surfacing fish detection, but also offers a useful technical basis for intelligent behavior monitoring, early warning, and data-driven environmental regulation in aquaculture systems.
Yang et al. (Sun,) studied this question.